For most of networking history, the physical layer was invisible to network engineers. You plugged in a cable, the link came up, and you moved on. The interesting work happened at layer 2 and above.

AI clusters broke this assumption along three dimensions simultaneously.

Bandwidth density. A DGX H100 node has eight ConnectX-7 NICs, each running at 400 Gbps. Eight ports x 400 Gbps = 3.2 Tbps per node of total network bandwidth. The switch that aggregates 32 DGX nodes must handle 32 x 400 Gbps = 12.8 Tbps on its downlink side alone -- and it must do this from a one-rack-unit chassis. Delivering that bandwidth density requires precision engineering at the photon level. There is no headroom for mediocre cabling.

Distance sensitivity. AllReduce in a DGX SuperPOD traverses four hops: NIC -> leaf -> spine -> leaf -> NIC. Each hop crosses a cable. At 400 Gbps with PAM-4 modulation, that cable must maintain single-photon-level signal integrity across every lane, simultaneously, with sub-nanosecond timing precision. A damaged cable that would have been tolerable at 10 Gbps destroys the signal at 400 Gbps. Physical quality standards that were aspirational in enterprise networking are mandatory minimums in AI clusters.

Thermal coupling. A single 64-port 800G leaf switch approaches 1,500 watts of power -- roughly 1,000 watts from the optics alone. Multiply by eight leaf switches in a DGX BasePOD and the switching layer alone consumes 12 kilowatts. The cooling infrastructure needed to support AI clusters is driven significantly by the optics inside the switches, not just the GPUs. The physical layer is an energy system, not just a signal system.

These three pressures are why the optics industry is reinventing itself around AI cluster requirements. Speed is doubling faster than transistor density. Form factors are being redesigned from scratch. The DSP -- the chip inside every transceiver that decodes the signal -- is migrating from pluggable modules into the switch ASIC itself.

Understanding this trajectory is not academic. The cables you are specifying today, the switch ports you are choosing, and the cabling infrastructure you are laying -- these decisions have a lifetime measured in years during which the cluster will scale through multiple optics generations.